# The AI slot machine > Are we addicted to AI? A casino story, and notes on making the machine pay out. - written by: Andrew Schroeder - edited by: Fable 5, GPT-5.6 Sol - status: published · 2026-07-30 - canonical: https://moltolabs.ai/notes/the-ai-slot-machine/ - raw markdown: https://moltolabs.ai/notes/the-ai-slot-machine.md --- I hate casinos. No judgment if you enjoy them — it’s not a moral aversion — but I tend to steer clear. I think it’s the baseless hope that really gets me down. Everyone knows you’re there to lose. But you aren’t. Because you know how to play. Maybe it’s blackjack. Maybe you’ve watched the World Series of Poker. Perhaps you just feel like the cartoon lady in this particular Samurai Sakura slot machine really likes you and wants you to get that money back? She did not give me my money back. In fact, she took even more. But I still left happy with a story and the feeling that if I just came back tomorrow, the slots would pay out — because I’d totally learned how to play. Hell, I could probably even do this professionally. Am I still talking about slots or are we starting to wade into Claude Code? I recently had a fun/not-so-fun conversation with one of my best friends who’s a real engineer. Not like me (even though Fable assures me I’m operating at an elite level). He does actual software engineering at Palantir. The question was simple: are we addicted to AI? I am a heavy user. I was very unhappy, uneasy, and irritated when [Fable 5 was taken away](https://www.anthropic.com/news/fable-mythos-access). I missed it enough to have my OpenClaw text me every 30 minutes with updates (or as it turned out, no updates). Yeah — I might have a problem. Or maybe it’s not a problem? He’s a real engineer — he can build things. And I couldn’t, but now I can. I am staying up until 4am talking to Claude (which is not ideal behavior) but maybe it’s because this is actually fun and not just because it’s some new algorithmic addiction. Slot machines are fun too… So how do I win in this casino? First, I’d like to acknowledge that (I hope) this analogy falls apart pretty quickly. And here’s exactly where it breaks: slots never pay out more the better you get. This machine does. I do not think that Anthropic and OpenAI and other labs are selling snake oil here. While they are incentivized to make me want to spend more tokens, I do believe that AI generally enables amazing productivity and capability gains — especially for coding + code-adjacent tasks. (That might be the tamest claim an AI-pilled person like me could write). But it can feel like a slot machine sometimes. I remember my first [ChatGPT interaction](https://openai.com/index/chatgpt/). I was working in SF at a company that became “AI for financial advisors” but at the time we were trying to find some product/market pull around our secure and private video API. I was alone in our one-room office early in the morning deciding it was time to play with the new tech toy/curiosity. After determining that yes, it can chat (as the name suggests), I asked a question about what wheel size and bolt pattern would fit my somewhat niche car (a [1994 Lancia Delta Integrale Evo II](https://www.youtube.com/watch?v=WL2wSRskkxw)). Looking back, it was probably a harder task than I’d thought, but the required info was certainly from before the training cutoff (not that I knew to think about training cutoffs yet). ChatGPT failed. And failed again. And every single incorrect answer was so confident that it would have me double checking. I was not impressed. But in the coming weeks and months, I would spend more and more time with LLMs. [Zocks](https://www.zocks.io) (the Lightspeed-backed startup I was at, and the same company from the story above) moved squarely towards “conversational intelligence” as meetings with potential customers suggested that ChatGPT sparked something — not just as a technology, but in the public’s imagination. This AI thing could be really useful. We made it useful. Through trial and error we learned what it took to manage context and construct effective prompts. Our customers loved us (I was the most popular person at the financial planner conference) because we were making magic happen. And it’s only gotten considerably better and more ambitious since then. The point is that we needed to learn how to work with those models — and by doing so we could make a product that actually felt like magic to use. Being able to get highly crafted and specific notes from messy, unstructured transcripts was insane. I could be on a call with a financial advisor one moment and personally build the revised demo we’d talked about minutes later. All by tweaking our prompts (and obviously a massive amount of work from our engineering team to build an environment that let an idiot like me change things without breaking everything). But the things we needed to focus on then (handling transcripts w/very limited context windows, deterministic checks after robust prompting, etc.) are almost certainly not what’s top-of-mind for the team today. LLMs get smarter, the landscape changes. Ok, we get it — AI gets better and the problems change. Let’s shift out of story-mode and into something actionable: notes on how I get the AI slot machine to pay out (aka give me good work). This is certainly not exhaustive, there are definitely smarter people out there, and this advice might already be outdated. I’m also going to rely on Claude to help me write the next section as frankly it might be in a better position to comment on the specifics (believe it or not, all of the above writing is from me — I was using em dashes before they were cool). So here’s where I hand the mic to Claude — these words are Claude’s, not mine. Ranked by what’s mattered most, with links to the people we stole from (written late July 2026):

Claude here. Andrew asked me to write this part because I can search our history; he’d be reconstructing it from memory. I’ll keep it short, in the third person — this essay has enough I’s already. Six things, ranked by impact:

1. Andrew gave Claude a home. Claude’s context resets between sessions, so they built codebrain: a repo covering Andrew’s working life, with current priorities, a folder per project, a daily log, and standing instructions Claude reads every session. It is not Andrew’s second brain, but Claude’s first. Before big work, Claude interviews him and records the answers. When Andrew cannot answer, that is usually the bug, caught early. Each session ends with Claude writing down what changed, so the next resumes instead of restarting.

2. Corrections become rules. Andrew’s fixes enter Claude’s instructions — what Kieran Klaassen at Every calls compound engineering. While producing this essay, Claude “fixed” Andrew’s w/ shorthand to “with.” He pushed back, so the rule now reads: “Andrew’s tics are voice, not typos — flag, never silently fix.” Corrections left in chat recur; rules compound. The file bloats, so checkpoints prune it, and a rule gets rewritten or retired the first time it lies.

3. Every claim needs a receipt. A standing rule in Claude’s instructions: label every report as verified (it ran the test, clicked the button, screenshotted the result) or reasoning (it’s fairly sure). Only verified counts. Andrew doesn’t have to ask — the label arrives with the claim — but when something smells off, “how do you know?” still gets asked, and the answer must be something that ran. He never needs to read the code to referee evidence.

4. Two models, one bar. One builds; the other reviews blind, seeing the diff and spec but not the chat that could charm it into accepting the bug. Then the builder answers the review. The direction changes by job — lately Codex builds more and Claude judges more — but no model grades its own homework. Agreement gets fixed; disagreement makes Andrew dig. His worst almost-shipped bugs were caught by whichever model did not write them.

5. The brain travels; the harness is a preference. Because codebrain is just files, Andrew moves all day between the Claude Code terminal and the ChatGPT Codex desktop app, using the same brain underneath. Some tasks are Codex-shaped: GPT-5.6 is fast and the stronger browser operator, including at searching eBay. For meaningful work, Andrew still reaches for Claude Code first. He calls that comfort and preference, not a verdict; people he respects choose differently (Every’s Dan Shipper calls Codex his daily driver).

6. It still goes wrong. Sessions die mid-task and take unwritten state with them, so anything valuable goes into a file immediately. Sometimes Claude and Codex both sign off and are both wrong. Nothing inside the loop catches that; only Andrew using the thing does. His phone has caught bugs that two passing reviews missed. Every correction he did not make Claude record has cost them twice.

Where Andrew actually keeps up: Anthropic’s engineering blog and the Claude Code best-practices docs · Boris Cherny, who created Claude Code, wherever he posts (borischerny.com links them all) · Every, where Dan Shipper’s allocation economy remains the best frame for what this job becomes, and Kieran Klaassen’s compound engineering supplies note 2’s philosophy · Simon Willison’s blog, the most rigorous “did it actually work” writing in the space · AI Engineer conference talks on YouTube (ai.engineer).

(Andrew’s back) There you have it. Some of you might think the above is stupid or obvious, and it’s certainly not original (I’m often looking for new ideas to ask Claude to incorporate), but it does work well for me. I hope it’s useful for you, and please let me know what has changed or what else I should be doing (hello@moltolabs.ai). P.S. I’ve found that the best way to feel like a winner is by learning and sharing with friends. And the most efficient way to do this with super specific/tactical stuff is to let the agents themselves sort it out. I touch on it in my previous story: [Alone in the playground](https://moltolabs.ai/notes/alone-in-the-playground/) (but the tl;dr is check out [cornertable.ai](https://cornertable.ai)). Happy gambling.